3/4/26: Unlearning Sensitive Information from AI: Principles, Scopes, and Emerging Challenges

Описание: As AI systems are increasingly deployed in sensitive domains, the ability to remove unwanted or harmful information becomes essential. This talk examines machine unlearning from a principled perspective, addressing three fundamental questions: what constitutes sensitive information, how we can reliably determine whether it has been forgotten, and where such knowledge resides within modern models. I will discuss realistic unlearning scenarios beyond simple data deletion, introduce evaluation and calibration methods for large language models, and analyze the structural fragility of latent knowledge for unlearning. These insights also highlight emerging challenges in balancing effective forgetting with model utility, advancing more trustworthy and accountable AI systems.

Speaker: Jianing Zhu is a Postdoctoral Fellow in the Chandra Department of Electrical and Computer Engineering at the University of Texas at Austin (https://zfancy.github.io/), advised by Professor Atlas Wang. He received his PhD at Hong Kong Baptist University and was a visiting researcher at CMU and RIKEN AIP. His research interests lie in trustworthy machine learning to advance model robustness, reliability, and transparency in modern AI systems, with a recent focus on machine unlearning and agentic memory.

Website: https://zfancy.github.io